Beat_this AI Model Tracks Beats and Downbeats Without DBN Postprocessing

Beat_this is an open-source AI model developed by Xavriley and collaborators at CPJKU, introduced in the ISMIR 2024 paper 'Beat This! Accurate Beat Tracking Without DBN Postprocessing.' The model detects beat positions and downbeat boundaries in audio files by deliberately avoiding Dynamic Bayesian Network postprocessing and traditional meter or tempo constraints, achieving state-of-the-art F1 scores. Its architecture combines convolutions with transformers operating across frequency and time dimensions, and was trained on diverse datasets including solo instruments, classical music, and pieces with time signature changes. Three full-sized model variants are available at roughly 78 MB each, alongside a compact 8.1 MB version, making it adaptable for use cases ranging from music transcription to dataset annotation. However, the model performs less well on continuity metrics compared to postprocessing-based methods and acknowledges limitations with difficult or underrepresented music genres.